[v2] vma: Detect infinite loop in vma tree

Commit Message

Liam R. Howlett Oct. 31, 2024, 7:36 p.m. UTC

From: "Liam R. Howlett" <[e-mail]>

There have been no reported infinite loops in the tree, but checking the
detection of an infinite loop during validation is simple enough.  Add
the detection to the validate_mm() function so that error reports are
clear and don't just report stalls.

This does not protect against internal maple tree issues, but it does
detect too many vmas being returned from the tree.

The variance of +10 is to allow for the debugging output to be more useful for
nearly correct counts.  In the event of more than 10 over the map_count, the
count will be set to -1 for easier identification of a potential infinite loop.

Note that the mmap lock is held to ensure a consistent tree state during the
validation process.

Signed-off-by: Liam R. Howlett <[e-mail]>
Cc: Lorenzo Stoakes <[e-mail]>
Cc: Vlastimil Babka <[e-mail]>
Cc: Jann Horn <[e-mail]>
---
 mm/vma.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Andrew Morton Oct. 31, 2024, 8:48 p.m. UTC | #1

On Thu, 31 Oct 2024 15:36:08 -0400 "Liam R. Howlett" <[e-mail]> wrote:

> From: "Liam R. Howlett" <[e-mail]>
> 
> There have been no reported infinite loops in the tree, but checking the
> detection of an infinite loop during validation is simple enough.  Add
> the detection to the validate_mm() function so that error reports are
> clear and don't just report stalls.
> 
> This does not protect against internal maple tree issues, but it does
> detect too many vmas being returned from the tree.
> 
> The variance of +10 is to allow for the debugging output to be more useful for
> nearly correct counts.  In the event of more than 10 over the map_count, the
> count will be set to -1 for easier identification of a potential infinite loop.
> 
> Note that the mmap lock is held to ensure a consistent tree state during the
> validation process.
> 
> ...
>
> +++ b/mm/vma.c
> @@ -615,7 +615,10 @@ void validate_mm(struct mm_struct *mm)
>  			anon_vma_unlock_read(anon_vma);
>  		}
>  #endif
> -		i++;
> +		if (++i > mm->map_count + 10) {
> +			i = -1;
> +			break;
> +		}
>  	}
>  	if (i != mm->map_count) {
>  		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);

It might be helpful to tell readers what's going on here?

--- a/mm/vma.c~vma-detect-infinite-loop-in-vma-tree-fix
+++ a/mm/vma.c
@@ -615,6 +615,7 @@ void validate_mm(struct mm_struct *mm)
 			anon_vma_unlock_read(anon_vma);
 		}
 #endif
+		/* Check for a infinite loop */
 		if (++i > mm->map_count + 10) {
 			i = -1;
 			break;

Liam R. Howlett Oct. 31, 2024, 9:21 p.m. UTC | #2

* Andrew Morton <[e-mail]> [241031 16:48]:
> On Thu, 31 Oct 2024 15:36:08 -0400 "Liam R. Howlett" <[e-mail]> wrote:
> 
> > From: "Liam R. Howlett" <[e-mail]>
> > 
> > There have been no reported infinite loops in the tree, but checking the
> > detection of an infinite loop during validation is simple enough.  Add
> > the detection to the validate_mm() function so that error reports are
> > clear and don't just report stalls.
> > 
> > This does not protect against internal maple tree issues, but it does
> > detect too many vmas being returned from the tree.
> > 
> > The variance of +10 is to allow for the debugging output to be more useful for
> > nearly correct counts.  In the event of more than 10 over the map_count, the
> > count will be set to -1 for easier identification of a potential infinite loop.
> > 
> > Note that the mmap lock is held to ensure a consistent tree state during the
> > validation process.
> > 
> > ...
> >
> > +++ b/mm/vma.c
> > @@ -615,7 +615,10 @@ void validate_mm(struct mm_struct *mm)
> >  			anon_vma_unlock_read(anon_vma);
> >  		}
> >  #endif
> > -		i++;
> > +		if (++i > mm->map_count + 10) {
> > +			i = -1;
> > +			break;
> > +		}
> >  	}
> >  	if (i != mm->map_count) {
> >  		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
> 
> It might be helpful to tell readers what's going on here?

Sounds good.  I guess I should have waited longer for a v2.

> 
> --- a/mm/vma.c~vma-detect-infinite-loop-in-vma-tree-fix
> +++ a/mm/vma.c
> @@ -615,6 +615,7 @@ void validate_mm(struct mm_struct *mm)
>  			anon_vma_unlock_read(anon_vma);
>  		}
>  #endif
> +		/* Check for a infinite loop */
>  		if (++i > mm->map_count + 10) {
>  			i = -1;
>  			break;
> _
>

David Hildenbrand Oct. 31, 2024, 9:42 p.m. UTC | #3

On 31.10.24 20:36, Liam R. Howlett wrote:
> From: "Liam R. Howlett" <[e-mail]>
> 
> There have been no reported infinite loops in the tree, but checking the
> detection of an infinite loop during validation is simple enough.  Add
> the detection to the validate_mm() function so that error reports are
> clear and don't just report stalls.
> 
> This does not protect against internal maple tree issues, but it does
> detect too many vmas being returned from the tree.
> 
> The variance of +10 is to allow for the debugging output to be more useful for
> nearly correct counts.  In the event of more than 10 over the map_count, the
> count will be set to -1 for easier identification of a potential infinite loop.
> 
> Note that the mmap lock is held to ensure a consistent tree state during the
> validation process.
> 
> Signed-off-by: Liam R. Howlett <[e-mail]>
> Cc: Lorenzo Stoakes <[e-mail]>
> Cc: Vlastimil Babka <[e-mail]>
> Cc: Jann Horn <[e-mail]>
> ---
>   mm/vma.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/vma.c b/mm/vma.c
> index 68138e8c153e..283e6bc4884f 100644
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -615,7 +615,10 @@ void validate_mm(struct mm_struct *mm)
>   			anon_vma_unlock_read(anon_vma);
>   		}
>   #endif
> -		i++;
> +		if (++i > mm->map_count + 10) {
> +			i = -1;
> +			break;
> +		}
>   	}
>   	if (i != mm->map_count) {
>   		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);


Reviewed-by: David Hildenbrand <[e-mail]>

Vlastimil Babka Oct. 31, 2024, 9:44 p.m. UTC | #4

On 10/31/24 20:36, Liam R. Howlett wrote:
> From: "Liam R. Howlett" <[e-mail]>
> 
> There have been no reported infinite loops in the tree, but checking the
> detection of an infinite loop during validation is simple enough.  Add
> the detection to the validate_mm() function so that error reports are
> clear and don't just report stalls.
> 
> This does not protect against internal maple tree issues, but it does
> detect too many vmas being returned from the tree.
> 
> The variance of +10 is to allow for the debugging output to be more useful for
> nearly correct counts.  In the event of more than 10 over the map_count, the
> count will be set to -1 for easier identification of a potential infinite loop.
> 
> Note that the mmap lock is held to ensure a consistent tree state during the
> validation process.
> 
> Signed-off-by: Liam R. Howlett <[e-mail]>
> Cc: Lorenzo Stoakes <[e-mail]>
> Cc: Vlastimil Babka <[e-mail]>
> Cc: Jann Horn <[e-mail]>

Reviewed-by: Vlastimil Babka <[e-mail]>

> ---
>  mm/vma.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/vma.c b/mm/vma.c
> index 68138e8c153e..283e6bc4884f 100644
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -615,7 +615,10 @@ void validate_mm(struct mm_struct *mm)
>  			anon_vma_unlock_read(anon_vma);
>  		}
>  #endif
> -		i++;
> +		if (++i > mm->map_count + 10) {
> +			i = -1;
> +			break;
> +		}
>  	}
>  	if (i != mm->map_count) {
>  		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);

Lorenzo Stoakes Nov. 1, 2024, 10:54 a.m. UTC | #5

On Thu, Oct 31, 2024 at 03:36:08PM -0400, Liam R. Howlett wrote:
> From: "Liam R. Howlett" <[e-mail]>
>
> There have been no reported infinite loops in the tree, but checking the
> detection of an infinite loop during validation is simple enough.  Add
> the detection to the validate_mm() function so that error reports are
> clear and don't just report stalls.
>
> This does not protect against internal maple tree issues, but it does
> detect too many vmas being returned from the tree.
>
> The variance of +10 is to allow for the debugging output to be more useful for
> nearly correct counts.  In the event of more than 10 over the map_count, the
> count will be set to -1 for easier identification of a potential infinite loop.
>
> Note that the mmap lock is held to ensure a consistent tree state during the
> validation process.
>
> Signed-off-by: Liam R. Howlett <[e-mail]>

Reviewed-by: Lorenzo Stoakes <[e-mail]>

> Cc: Lorenzo Stoakes <[e-mail]>
> Cc: Vlastimil Babka <[e-mail]>
> Cc: Jann Horn <[e-mail]>
> ---
>  mm/vma.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/vma.c b/mm/vma.c
> index 68138e8c153e..283e6bc4884f 100644
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -615,7 +615,10 @@ void validate_mm(struct mm_struct *mm)
>  			anon_vma_unlock_read(anon_vma);
>  		}
>  #endif
> -		i++;
> +		if (++i > mm->map_count + 10) {
> +			i = -1;
> +			break;
> +		}
>  	}
>  	if (i != mm->map_count) {
>  		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
> --
> 2.43.0
>

Patch

diff --git a/mm/vma.c b/mm/vma.c
index 68138e8c153e..283e6bc4884f 100644
--- a/mm/vma.c
+++ b/mm/vma.c
@@ -615,7 +615,10 @@  void validate_mm(struct mm_struct *mm)
 			anon_vma_unlock_read(anon_vma);
 		}
 #endif
-		i++;
+		if (++i > mm->map_count + 10) {
+			i = -1;
+			break;
+		}
 	}
 	if (i != mm->map_count) {
 		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);